Skip to contentSkip to conversation

net: introduce skb_coalesce_rx_frag()

PATCH Superseded Jason Wang 2 participants 5 messages Oct 31, 2013 – Nov 1, 2013

Subsystems (4)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Jason Wang

author
#0 lore raw

PATCH v2 1/2 net: introduce skb_coalesce_rx_frag()

Sometimes we need to coalesce the rx frags to avoid frag list. One example is virtio-net driver which tries to use small frags for both MTU sized packet and GSO packet. So this patch introduce skb_coalesce_rx_frag() to do this.

Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Michael Dalton <[email protected]>
Cc Eric Dumazet <[email protected]>

---
Changes from V1:
- remove the useless off parameter.

---
 include/linux/skbuff.h |  3 +++
 net/core/skbuff.c      | 13 +++++++++++++
 2 files changed, 16 insertions(+)
include/linux/skbuff.h +3 -0
index 2c15497..fffaeaf 100644
@@ -1372,6 +1372,9 @@ static inline void skb_fill_page_desc(struct sk_buff *skb, int i,
 void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off,
 		     int size, unsigned int truesize);
 
+void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,
+			  unsigned int truesize);
+
 #define SKB_PAGE_ASSERT(skb) 	BUG_ON(skb_shinfo(skb)->nr_frags)
 #define SKB_FRAG_ASSERT(skb) 	BUG_ON(skb_has_frag_list(skb))
 #define SKB_LINEAR_ASSERT(skb)  BUG_ON(skb_is_nonlinear(skb))
net/core/skbuff.c +13 -0
index 0ab32fa..87670e1 100644
@@ -476,6 +476,19 @@ void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off,
 }
 EXPORT_SYMBOL(skb_add_rx_frag);
 
+void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,
+			  unsigned int truesize)
+{
+	skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
+
+	skb_frag_size_add(frag, size);
+	skb->len += size;
+	skb->data_len += size;
+	skb->truesize += truesize;
+	skb_frag_unref(skb, i);
+}
+EXPORT_SYMBOL(skb_coalesce_rx_frag);
+
 static void skb_drop_list(struct sk_buff **listp)
 {
 	kfree_skb_list(*listp);
Signature · 1 line
1.8.1.2
author
#1 lore raw

PATCH v2 2/2 virtio-net: coalesce rx frags when possible during rx

Commit 2613af0ed18a11d5c566a81f9a6510b73180660a (virtio_net: migrate mergeable rx buffers to page frag allocators) try to increase the payload/truesize for MTU-sized traffic. But this will introduce the extra overhead for GSO packets received because of the frag list. This commit tries to reduce this issue by coalesce the possible rx frags when possible during rx. Test result shows the about 15% improvement on full size GSO packet receiving (and even better than commit 2613af0ed18a11d5c566a81f9a6510b73180660a).

Before this commit:
./netperf -H 192.168.100.4
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.100.4 () port 0 AF_INET : demo

Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

87380 16384 16384 10.00 20303.87

After this commit:
./netperf -H 192.168.100.4
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.100.4 () port 0 AF_INET : demo

Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

87380 16384 16384 10.00 23841.26

Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Michael Dalton <[email protected]>
Cc Eric Dumazet <[email protected]>
---
 drivers/net/virtio_net.c | 14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)
drivers/net/virtio_net.c +10 -4
index 113ee93..5dc0de0 100644
@@ -305,7 +305,7 @@ static int receive_mergeable(struct receive_queue *rq, struct sk_buff *head_skb)
 	struct sk_buff *curr_skb = head_skb;
 	char *buf;
 	struct page *page;
-	int num_buf, len;
+	int num_buf, len, offset;
 
 	num_buf = hdr->mhdr.num_buffers;
 	while (--num_buf) {
@@ -342,9 +342,15 @@ static int receive_mergeable(struct receive_queue *rq, struct sk_buff *head_skb)
 			head_skb->truesize += MAX_PACKET_LEN;
 		}
 		page = virt_to_head_page(buf);
-		skb_add_rx_frag(curr_skb, num_skb_frags, page,
-				buf - (char *)page_address(page), len,
-				MAX_PACKET_LEN);
+		offset = buf - (char *)page_address(page);
+		if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
+			skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
+					     len, MAX_PACKET_LEN);
+		} else {
+			skb_add_rx_frag(curr_skb, num_skb_frags, page,
+					offset, len,
+					MAX_PACKET_LEN);
+		}
 		--rq->num;
 	}
 	return 0;
Signature · 1 line
1.8.1.2
Acked-by on v2 2/2
#2 lore raw

Re: [PATCH net-next V2 2/2] virtio-net: coalesce rx frags when possible during rx

Quoted 29 lines · Jason Wang

Commit 2613af0ed18a11d5c566a81f9a6510b73180660a (virtio_net: migrate mergeable rx buffers to page frag allocators) try to increase the payload/truesize for MTU-sized traffic. But this will introduce the extra overhead for GSO packets received because of the frag list. This commit tries to reduce this issue by coalesce the possible rx frags when possible during rx. Test result shows the about 15% improvement on full size GSO packet receiving (and even better than commit 2613af0ed18a11d5c566a81f9a6510b73180660a).

Before this commit:
./netperf -H 192.168.100.4
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.100.4 () port 0 AF_INET : demo

Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

87380 16384 16384 10.00 20303.87

After this commit:
./netperf -H 192.168.100.4
MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.100.4 () port 0 AF_INET : demo

Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

87380 16384 16384 10.00 23841.26

Quoted 7 lines
Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Michael Dalton <[email protected]>
Cc Eric Dumazet <[email protected]>

---

Excellent !

We now have 2 or 3 frags per skb, like tcp stack manages to do on output path.

Michael Dalton is also working on a autotuning patch, using an EWMA, so that the size of individual sg blocks can vary from 1500 to 4096, this might show even better throughput, we'll see.

#3 lore raw
In reply to Jason Wang
Quoted 42 lines · Jason Wang

Sometimes we need to coalesce the rx frags to avoid frag list. One example is virtio-net driver which tries to use small frags for both MTU sized packet and GSO packet. So this patch introduce skb_coalesce_rx_frag() to do this.

Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Michael Dalton <[email protected]>
Cc Eric Dumazet <[email protected]>

---
Changes from V1:
- remove the useless off parameter.

---
 include/linux/skbuff.h |  3 +++
 net/core/skbuff.c      | 13 +++++++++++++
 2 files changed, 16 insertions(+)
include/linux/skbuff.h +3 -0
index 2c15497..fffaeaf 100644
@@ -1372,6 +1372,9 @@ static inline void skb_fill_page_desc(struct sk_buff *skb, int i,
 void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off,
 		     int size, unsigned int truesize);
 
+void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,
+			  unsigned int truesize);
+
 #define SKB_PAGE_ASSERT(skb) 	BUG_ON(skb_shinfo(skb)->nr_frags)
 #define SKB_FRAG_ASSERT(skb) 	BUG_ON(skb_has_frag_list(skb))
 #define SKB_LINEAR_ASSERT(skb)  BUG_ON(skb_is_nonlinear(skb))
net/core/skbuff.c +2 -0
index 0ab32fa..87670e1 100644
@@ -476,6 +476,19 @@ void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off,
 }
 EXPORT_SYMBOL(skb_add_rx_frag);
 
+void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,
+			  unsigned int truesize)

net/core/skbuff.c void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off, · line 490

Quoted 10 lines
net/core/skbuff.c +8 -0
+{
+	skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
+
+	skb_frag_size_add(frag, size);
+	skb->len += size;
+	skb->data_len += size;
+	skb->truesize += truesize;
 
 
+	skb_frag_unref(skb, i);

This unref is not logical, or should at least be

__skb_frag_unref(frag);

But I do think this is best done in the caller.

In virtio_net this would be a :

put_page(page);

In tcp stack we do almost the same, but we take the reference on the page if we could not coalesce with prio frag, instead of doing a get and put in the other case.

        if (can_coalesce) {
                skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
        } else {
                get_page(page);
                skb_fill_page_desc(skb, i, page, offset, copy);
        }
Quoted 6 lines
net/core/skbuff.c +3 -0
+}
+EXPORT_SYMBOL(skb_coalesce_rx_frag);
+
 static void skb_drop_list(struct sk_buff **listp)

{
kfree_skb_list(*listp);

author
#4 lore raw
Quoted 66 lines · Eric Dumazet
Quoted 41 lines · Jason Wang

Sometimes we need to coalesce the rx frags to avoid frag list. One example is virtio-net driver which tries to use small frags for both MTU sized packet and GSO packet. So this patch introduce skb_coalesce_rx_frag() to do this.

Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Michael Dalton <[email protected]>
Cc Eric Dumazet <[email protected]>

---
Changes from V1:
- remove the useless off parameter.

---
 include/linux/skbuff.h |  3 +++
 net/core/skbuff.c      | 13 +++++++++++++
 2 files changed, 16 insertions(+)
include/linux/skbuff.h +3 -0
index 2c15497..fffaeaf 100644
@@ -1372,6 +1372,9 @@ static inline void skb_fill_page_desc(struct sk_buff *skb, int i,
 void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off,
 		     int size, unsigned int truesize);
 
+void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,
+			  unsigned int truesize);
+
 #define SKB_PAGE_ASSERT(skb) 	BUG_ON(skb_shinfo(skb)->nr_frags)
 #define SKB_FRAG_ASSERT(skb) 	BUG_ON(skb_has_frag_list(skb))
 #define SKB_LINEAR_ASSERT(skb)  BUG_ON(skb_is_nonlinear(skb))
net/core/skbuff.c +1 -0
index 0ab32fa..87670e1 100644
@@ -476,6 +476,19 @@ void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off,
 }
 EXPORT_SYMBOL(skb_add_rx_frag);
 
+void skb_coalesce_rx_frag(struct sk_buff *skb, int i, int size,

net/core/skbuff.c void skb_add_rx_frag(struct sk_buff *skb, int i, struct page *page, int off, · line 489

Quoted 10 lines
net/core/skbuff.c +9 -0
+			  unsigned int truesize)
+{
+	skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
+
+	skb_frag_size_add(frag, size);
+	skb->len += size;
+	skb->data_len += size;
+	skb->truesize += truesize;
 
+	skb_frag_unref(skb, i);

This unref is not logical, or should at least be

__skb_frag_unref(frag);

But I do think this is best done in the caller.

In virtio_net this would be a :

put_page(page);

In tcp stack we do almost the same, but we take the reference on the page if we could not coalesce with prio frag, instead of doing a get and put in the other case.

Quoted 7 lines
        if (can_coalesce) {
                skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
        } else {
                get_page(page);
                skb_fill_page_desc(skb, i, page, offset, copy);
        }

Ok, get it. Will do a put_page() in V3.

Thanks

Quoted 11 lines
Quoted 6 lines
net/core/skbuff.c +3 -0
+}
+EXPORT_SYMBOL(skb_coalesce_rx_frag);
+
 static void skb_drop_list(struct sk_buff **listp)
 {

kfree_skb_list(*listp);

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html